A Data-Driven Framework for Optimized System Maintenance
Ubale, Salim, Remenyte-Prescott, Rasa, Grant, David, Stuart, Alastair, Hague, Adam, Saleh, Ali · HAL (Le Centre pour la Communication Scientifique Directe) · 2019
Conventional process controllers (such as proportional integral derivative controllers and model predictive controllers) are simple and effective once they have been calibrated for a given system. However, it is difficult and costly to re-tune these controllers if the system deviates from its normal conditions and starts to deteriorate. Recently, reinforcement learning has shown a significant improvement in learning process control policies through direct interaction with a system, without the need of a process model or the system characteristics, as it learns the optimal control by interacting with the environment directly. However, developing such a black-box system is a challenge when the system is complex and it may not be possible to capture the complete dynamics of the system with just a single reinforcement learning agent. Therefore, in this paper, we propose a simple architecture that does not replace the conventional proportional integral derivative controllers but instead augments the control input to the system with a reinforcement learning agent. The agent adds a correction factor to the output provided by the conventional controller to maintain optimal process control even when the system is not operating under its normal condition.